Vocap

Situation overview. A public debate has opened over whether the largest AI companies should slow the development of their most powerful AI systems. Dario Amodei, CEO of Anthropic (maker of Claude), has called for slowing the pace so independent safety testing can keep up. Sam Altman, CEO of OpenAI, and several other technology leaders quickly expressed support. President Trump and some of his advisers strongly disagree, arguing that slowing U.S. development could give China an advantage.

The most alarming headline came a few days earlier. Evan Hubinger, Anthropic’s alignment-science lead — essentially someone responsible for making sure increasingly powerful AI behaves as humans intend — said he believes there is greater than a 10% chance AI could kill all humans within the next decade. Another Anthropic researcher, Jacob Coxon, resigned that same week, saying that Anthropic and OpenAI were not acting responsibly enough given the risks. These are personal judgments about far more powerful future AI systems, not today’s AI and not a settled forecast. But they help explain why the debate has suddenly become much more urgent.

Our view. AI’s economic impact is being limited less by intelligence than by implementation. Today’s models can already do far more than most businesses have managed to put into production. The constraint is delivery, which means connecting AI to the systems a company actually runs on, making it reliable enough to trust with consequential work, and reorganizing workflows around it. Intelligence has run well ahead of integration, and integration is where our companies operate. A slower frontier does not slow that work down. It gives it room.

What Happened

Dario Amodei published an essay, “We Must Pace the Frontier,” arguing that AI is improving faster than anyone can reliably test or control it. The CEO of Anthropic is particularly concerned about future systems that could operate with less human supervision, assist with cyberattacks, or even help design more powerful versions of AI.
He did not ask anyone to stop. He proposed independent reviewers working inside the leading AI companies with real access, safety checks that a system must clear before it advances, and eventually agreements among governments.
The industry agreed almost immediately. Sam Altman said OpenAI would use embedded reviewers too. Demis Hassabis of Google DeepMind called the direction correct. Elon Musk posted “Dario is right.” Microsoft welcomed the idea without committing to it. Meta said nothing.
Washington pushed back hard. Former White House AI adviser David Sacks called the argument fearmongering designed to lock in today’s leaders. On September 14, President Trump phoned Nvidia CEO Jensen Huang live on stage at a technology conference and called AI safety concerns “a hoax.” China’s foreign ministry called pause proposals an attempt to hold rivals back.
The financial backdrop is relevant. The Financial Times reported on September 13 that Anthropic generated $11.5 billion of Q2 revenue, roughly 14x the prior year, and was profitable on an adjusted operating basis. The company is also preparing for an IPO that could value it at $2 trillion or more. For an established leader like Anthropic, slowing the race to build ever-more-powerful models could also have financial benefits. It could reduce enormous development costs, protect its current competitive position, and make it harder for rivals to close the gap. In other words, Amodei’s safety concerns may be genuine, but the approach he is advocating could also serve Anthropic’s economic interests.

Why the Disagreement Is So Sharp

This is not a left versus right fight. Two other divides explain the heat better:

Safety against speed. One camp believes competition pushes companies to release systems before anyone fully understands them, and that a modest delay buys time for testing. The other believes the worst scenarios are speculative and that slowing American companies hands ground to China.
Incumbents against everyone else. Rules built around audits and certifications are easier for large, established companies to satisfy. Critics argue that this could protect the incumbents by raising the cost of competing. Supporters argue the safety risks are real regardless of who benefits. Both can be true at once.

The deeper problem is that the two camps are not weighing the same risk differently but disagreeing about whether it exists, which is why neither finds the other’s argument responsive. The case for speed quietly assumes what the case for caution denies: that whoever arrives first will be able to control what they have built.

What It Could Mean for AI Investing

Markets read “slower AI” as bad news for the companies selling chips, power and data centers, on the logic that fewer giant training projects means less demand. Other analysts argue that is the wrong conclusion, and we agree.

The shortage is in delivery, not intelligence. The current race repeatedly makes new frontier models obsolete before the economy has finished absorbing it. Each new model release forces software companies to retest products, rebuild connections and re-explain themselves, which is time not spent getting AI into production.
Computing power gets redirected, not switched off. Computing capacity is the scarce input the entire industry is competing for. The largest technology companies are all chasing the same chips, power and land. If the biggest AI training projects use less of that capacity, it does not sit idle. More of it can be used to run AI applications for customers, which the industry calls “inference.”
Safety turns into a competitive asset. If operating at the leading edge starts to require audits and certifications, the scarce advantage shifts from having the smartest model to having permission to run one. That favors whoever already paid for that infrastructure and starts to make a frontier AI lab look more like a regulated utility than a software company.

What This Changes for Our Portfolio

Our companies close that delivery gap for a living. Their advantage comes from understanding a customer’s operations, fitting into daily work, and producing a dependable result. This is the thesis we published earlier this year in our SaaSpocalypse pieces: the intelligence layer is becoming a commodity, and the execution layer is the prize.

Every reset is time taken from that work. What damages a specialized software company is not a smarter general model. It is having to be rebuilt instead of compound. Fewer resets mean more time accumulating what actually defends a business: domain knowledge, customer-specific data, and embedded workflow. Our returns depend on hospitals, insurers, manufacturers and logistics operators putting AI to work, not on any one lab producing the next breakthrough.
Two risks we are watching. First, large AI companies with more breathing room and better economics have more capacity to move into the software markets our portfolio serves. Second, if safety becomes a requirement to operate, that burden travels downstream. We expect enterprise buyers to start asking our companies which models they use, how they test them and how results can be audited, and we expect that to reach board agendas during 2027.

What Would Prove Us Wrong

Our view changes if the leading AI companies never actually slow down; if freed-up computing capacity does not move toward customer applications, making this real demand destruction rather than reallocation; if new rules raise costs without creating real advantage for those who comply; or if the largest AI companies use their improved economics to compete directly against the businesses we back.

Bottom Line

This debate is not a reason to turn negative on AI. It is a shift where money gets spent inside the industry, less on the race to build the next model and more on the far larger job of putting the models we already have to work. That is the job our companies do. Our investment criteria and pace of investment are unchanged.

One caveat worth stating: every participant here has incentives, including us. Amodei can believe the safety problem is urgent and still be proposing a structure that turns his company’s biggest historical expense into its strongest asset, weeks before an IPO. We would rather give you the mechanics than the verdict.